Learning Video Exposure Correction: Large-Scale Real-World Dataset and Method.

Abstract

Exposure correction aims to enhance visual data degraded by improper exposures, thereby improving perceptual quality. While substantial progress has been achieved in single-image exposure correction, its extension to videos remains largely underexplored. Applying frame- wise correction methods to videos often results in flickering artifacts and degraded visual coherence. Progress in video exposure correction has also been hindered by the lack of a large-scale paired benchmark. To address this limitation, we introduce a real-world paired video exposure correction dataset that covers both underexposed and overexposed dynamic scenes. The dataset comprises 36 K spatially aligned frame pairs acquired using a dual-camera acquisition system equipped with a beam splitter, which enables synchronized recording of ill-exposed and well-exposed videos. Based on this dataset, we develop an end-to-end dual-stream Retinex framework that jointly handles underexposure and overexposure through exposure-aware temporal modeling. To mitigate exposure-induced interference during temporal alignment, our framework aggregates the amplitude spectra of neighboring frames into a shared exposure representation while retaining phase-derived structural information for cross-frame correspondence. The framework further reinterprets residual inter-frame discrepancies as pseudo- gradient cues for inverted attention, enabling temporal fusion to preserve complementary information while suppressing unreliable content. Finally, a learned restoration module combines the disentangled reflectance and dual-stream illumination components to produce temporally consistent exposure-corrected frames. The extensive experiments based on various evaluations and user studies demonstrate the significance of our dataset and the effectiveness of our method. Our dataset is available at https://dx.doi.org/10.21227/w9yn-y292.